Reciprocal style and information transfer between historical Istanbul Pervititch Maps and satellite views using machine learning

Abstract

Historical maps contain significant data on the cultural, social, and urban character of cities. However, most historical maps utilize specific notation methods that differ from those commonly used today and converting these maps to more recent formats can be highly labor-intensive. This study is intended to demonstrate how a machine learning (ML) technique can be used to transform old maps of Istanbul into spatial data that simulates modern satellite views (SVs) through a reciprocal map conversion framework. With this aim, the Istanbul Pervititch Maps (IPMs) made by Jacques Pervititch in 1922-1945 and current SVs were used to test and evaluate the proposed framework. The study consists of a style and information transfer in two stages: (i) from IPMs to SVs, and (ii) from SVs to IPMs using CycleGAN (a type of generative adversarial network). The initial results indicate that the proposed framework can transfer attributes such as green areas, construction techniques/materials, and labels/tags.